AI编码
dev.to ·
@Ekene Eze (Kenny)
产品文档的未来：当 agent 成为主要读者
文章指出，未来客户可能由编码助手代为阅读文档并生成代码，因此文档准确性成为持续责任。文中引用 Mintlify 2026 State of Knowledge 报告：其站点 8 月记录 2.57 亿次 agent 请求，对比 1.31 亿次人类页面加载，但仅 7% 受访公司让所有知识面可被 agent 读取。作者提醒，过时示例通过 Markdown 或 MCP 被检索，只会让错误更容易被重复。
AI推荐理由 引用 Mintlify 2026 报告数据（8 月 agent 请求 2.57 亿次 vs 人类页面加载 1.31 亿次，仅 7% 受访公司使全部知识面可被 agent 读取），讨论文档面向 agent 消费后的准确性问题。数据来自厂商报告，需注意其立场，但议题对 AI 应用与开发工具生态有实际意义。
64
正文
Your next customer might integrate your product without ever visiting your documentation site. Their coding assistant will read it for them, choose an approach, and generate the code.
What happens when those instructions are outdated?
Mintlify’s 2026 State of Knowledge Report highlights the scale of this shift. In August, its sites recorded 257 million agent requests compared with 131 million human page loads. Yet only 7% of surveyed companies had made every knowledge surface readable by agents.
Making that knowledge accessible is essential. Keeping it accurate is an ongoing responsibility.
An outdated example served through Markdown or MCP still teaches the wrong approach. Making it easier to retrieve can make the mistake easier to repeat.
Consider that you shipped a change to your API authentication flow. Engineering ships it, and someone in DevRel updates the API reference. But the old approach remains in a quickstart, a support article, and an integration example.
Which version will a customer’s agent find?
At LI.FI, Netlify, Flutterwave, and Shardeum, I experienced how much documentation work depended on someone remembering those connections.
As products grow, that becomes harder to sustain.
I believe keeping product knowledge aligned should become a routine part of shipping software.
Every meaningful product change should prompt a review of the instructions it affects. Proposed updates should show the evidence behind them.
People should retain control over what gets published, especially when release timing, version support, or product intent requires judgment.
That is the thinking behind Thally.
Thally connects product repositories with the places teams explain their products. It identifies affected content and opens evidence-backed pull requests with proposed updates for the team to review.
Its documentation platform serves people and agents from the same source. The ongoing responsibility is keeping that source aligned with the product.
The question I want teams to ask after a release is simple:
Can someone following our published instructions still succeed?
That should be part of deciding whether the release is complete.
thally.io